Re: [PATCH v6 18/18] RDMA/mlx5: Ask P2PDMA whether ATS takes a direct peer-to-peer route

[Jason Gunthorpe]

On Thu, Sep 17, 2026 at 04:16:57PM +0200, Thomas Hellström wrote:
> On Thu, 2026-09-17 at 11:02 -0300, Jason Gunthorpe wrote:
> > On Thu, Sep 17, 2026 at 03:56:12PM +0200, Thomas Hellström wrote:
> > 
> > > That is, a flag to tell the topology check that some transactions
> > > *will* take the host-bridge path due to IOVA being used, and that
> > > the
> > > computations including pci_p2pdma_distance() need to check whether
> > > that
> > > is possible (checking whitelist etc.) and return the corresponding
> > > THRU_HOST_BRIDGE mapping type. Translated transactions taking a
> > > short-
> > > cut using the bus-address would then be hidden from the driver.
> > 
> > I imagined it would work by the driver passing in what modes it
> > supports and Intel GPU would simply pass in ATS only.
> > 
> > Even something like mlx has variation, only some operations can opt
> > into ATS, many others are never-ATS.
> > 
> > Driver should not make statements like " *will* take the host-bridge
> > path" since it has no idea.
> 
> Yes, agreed, but anyway DMA-BUF would then, as part of the
> dynamic_attach negotiation have an interface that takes an ATS only
> flag and use it to determine whether P2P is possible at all, and use
> them at mapping time to set up the correct SG list (IOVA or BUS
> address).

Yeah, where I was going with mapping type was to have the core code
"DMA buf PCI bus mapping type" perform the pci p2p operations,
(especially the calc distance that has been open coded in GPU
drivers!!) and I imagined it would take in information from the
exporter and importer about things like ATS capabilities.

There is worse coming here soon people are working on multi-path PCI
and we even already have such systems. Multi-path PCI means you have
to select the best quality path from a NxM matrix of options and then
inform both sides which path you picked.

> Similarly the dma_map_page() needs to know these flags for p2pdma pages
> so that it never sets up a BUS_ADDRESS to an ATS-enabled device that
> didn't explicitly ask for it?

You don't use dma_map_page() with P2P, use the new interface and then
the details are visible and the mapper can do the right logic.

> And this all (and in addition the previous discussion around pcie
> subsystems calls from within dma-buf, and data-types for mapping
> information) makes me think that abstracting fast dma-buf interconnects
> along with our previous discussions is a good thing. Then interconnect
> details and cross-subsystem dependencies can be hidden in the
> interconnect implementations, suitably placed in the tree.

Right exactly, I guess Christain will be more comfortable if we have a
dma-mapping-type-pci.c that has all of this and only PCI drivers use
that code so it makes logical layering sense with the needs of the
embedded non-PCI drivers..

Jason